How lawyers should think about AI
Legal work is dense with text — contracts, filings, case law, correspondence — and AI is fastest exactly where volume of text is the bottleneck. But the profession’s duties of confidentiality, competence, and candour raise the stakes of a wrong answer, and general language models are notorious for fabricating citations that look real. The workable posture is to use AI to accelerate drafting, review, and research while treating every output as an unverified draft that you, the qualified lawyer, are accountable for. This guide maps where AI helps and the boundaries you cannot cross.
Contract review and analysis
AI is strong at first-pass contract triage. Provide a contract and ask it to summarise key terms, flag unusual or missing clauses, compare against a standard template, or extract obligations and dates into a table. This turns hours of reading into a focused review of the points the model surfaced. The discipline is to verify each flag against the actual text — the model can misread, miss context, or misstate a clause — and to never let an AI summary substitute for reading the operative provisions that matter to your client.
Legal research with RAG
General models should never be trusted to recall case law from memory, because they invent citations. The safer pattern is retrieval-augmented generation (RAG): the tool searches a real, authoritative legal database and answers only from the documents it retrieved, with links to the sources. Even then, open and read the primary authority yourself before relying on it. Understanding how models consume long documents — see What Is a Token in AI? — helps you work within their limits when feeding in lengthy materials.
Drafting
AI accelerates first drafts of routine documents: standard clauses, client letters, due-diligence checklists, and plain-English explanations of complex terms. Give the model clear instructions, the relevant facts, the tone, and the jurisdiction, and treat the result as a starting point to be edited heavily. It is excellent at structure and boilerplate, and unreliable on anything requiring current, jurisdiction-specific legal accuracy — which is most of what matters.
Ethics, confidentiality, and the hard line
Three duties govern every use. Confidentiality and privilege: do not put client information into tools that may store or train on it; use only vetted platforms with appropriate data protections. Competence: understand what the tool can and cannot do, and supervise it. Candour and verification: never file or send AI output you have not checked, especially citations — courts have already sanctioned lawyers for fabricated cases. The final boundary is responsibility itself: AI carries none, you carry all. Used as a fast, supervised assistant within these limits, it makes legal work quicker without putting your clients, your filings, or your licence at risk.